Physix FrontierCompute Surplus Is a False Premise
Core Judgment: Computing power surplus is not a total volume issue, but one of structural mismatch and scheduling efficiency. Meta selling computing power and Anthropic buying it—viewed together, these two events precisely indicate that the computing power market is undergoing a "de-bubbling" selection process.
From an engineering perspective, computing power has never been a binary question of "enough or not." Having worked on embedded systems for years, I've seen many debates about "computing power surplus"—for example, some say smartphone chip performance is excessive, but put it on a drone flight controller, and the real-time requirements for IMU data fusion cause CPU utilization to spike above 90% in minutes. So-called "surplus" just means you haven't found the right application scenario that is latency-sensitive and has hard throughput constraints.
If you only look at the surface, Meta selling "excess AI computing power" does seem to lead to the conclusion that "computing power is starting to become surplus." But looking at the timeline, Meta made large-scale purchases of H100s in 2022, during the early stages of the large model arms race, when everyone was frantically hoarding inventory. Now that Meta wants to sell, it's more about asset allocation optimization—some cards have low utilization, others are being replaced by new architectures, so it makes sense to cash out. This follows the same logic as Bitcoin mining rig trading back then: miners don't lack computing power needs; rather, computing power depreciates too fast, requiring constant equipment upgrades.
Anthropic signing a $19 billion, 20-year data center lease is the real signal. This money isn't buying graphics cards; it's buying long-term stable computing supply, bundled with entire data center infrastructure. What does a 20-year lease imply? It implies they believe computing demand will only grow, not shrink, over the next 20 years. If computing power were truly in surplus, they could easily rent cheaply on the spot market; there would be no need to lock in cash flow for 20 years.
Viewing these two events together, a more reasonable explanation is: the computing power market is shifting from "barbaric hoarding" to "refined operations." Meta is selling "redundant assets," while Anthropic is buying "deterministic capacity." It's like a factory where someone sells idle CNC machines while another pays big money for long-term power contracts—the focus is completely different.
As a flight control engineer, I'm used to breaking problems down into "real-time requirements" and "average throughput." The structure of computing demand for AI training versus inference differs greatly. Training requires massive parallel computation, insensitive to latency but demanding high throughput; inference, especially edge inference, is latency-sensitive and has strict constraints on computing density and power consumption. The computing power Meta sold off is likely general-purpose GPUs targeted at training scenarios, whereas what Anthropic needs is a stable computing pool supporting inference clusters. These are fundamentally different things.
Another easily overlooked detail is the actual utilization rate of computing power. Current GPU clusters, especially in distributed training scenarios, do not have high actual computing utilization. I've seen teams using 8,000-card clusters to train models, but communication overhead accounts for 40% of the time, leaving computing utilization at only 60%. If computing power were truly in surplus, they should first optimize scheduling algorithms to improve utilization, not directly sell cards. Meta's choice to sell cards suggests they consider the cost of "optimizing scheduling" higher than "selling cards for cash," or they have more urgent internal cash flow needs.
From an engineering feasibility standpoint, the proposition of "computing power surplus" is hard to sustain. Because the speed of progress in chip manufacturing processes and architectures currently cannot keep up with the growth speed of model parameters and inference loads. GPT-4 has 1.8 trillion parameters; GPT-5 might be larger. Even if computing power doubles every year, it may not catch up with model scale growth. Not to mention scenarios like autonomous driving, robotics, and industrial AI, where the demand for real-time computing is infinite.
So my advice is: Don't get misled by macro narratives like "computing power surplus." If you work in AI infrastructure, pay more attention to the micro-structure of computing distribution—which scenarios need low latency, which need high throughput, and which are power-sensitive. Treat computing power as a resource scheduling problem, not a total inventory problem. What you should really worry about isn't having too much computing power, but whether your computing power is being used in the right places.
Original Link: https://www.tmtpost.com/8061099.html